Top positive review
5.0 out of 5 starsCurrently the world's best single CPU for rendering, and some other compute-heavy tasks
Reviewed in the United States on November 6, 2018
Currently, this CPU is hands down objectively the best at 3D rendering in a single socket (by the metric of time taken to render 3D scenes).
In blender cycles, I've had this CPU outpace my dual 1080ti at rendering, 950 vs 693 seconds in a scene I've been using to benchmark.
I also do a lot of data processing.
Encoding (ffmpeg): depending on the codec used, it can be pretty good, or pretty meh, in the worst cases getting slightly outpaced by the 18 core i9 7980XE in h265. h264 I found was not quite as favorable for the intel cpus though, but something to note is that neither of those codecs actually fully loaded the cpu, 2 of the dies were typically idle, thus the cpu was being treated like a 16-core, and one of the dies was barely loaded, resulting in severe under utilization for the AMD cpu that pushed its boost clocks down, while the intel cpu was able to make use of much higher boosts, since at least the number of threads was much closer to its actual core count. Parallel encoding however swings things back in the favor of the 2990WX though, where it can now fully utilize all the cores. Occasionally some under utilization would happen due to memory bandwidth constraints in some portions of my reference task.
NLP Text Preprocessing (Proprietary w/ Python, Numpy): This particular bit of code is unfortunately a custom piece written for my job, so not necessarily directly translatable to what other people may be doing, but running 64 instances of the program in parallel didnt scale how I expected... because my nvme drives couldn't keep up with the IO demands, resulting in CPU under utilization, so not the CPU's fault. I'll need to set up a larger scratch disk array to properly test this.
AVX2 deep learning: ok, the 1080tis win out for this workload, but against other cpus the 2990WX can still hold its own... until you bring in the question of AVX512, which is an interesting topic. Sure, AVX512 might be a bit easier to code for than CUDA... in theory, since its still all on the CPU. However AVX512 is not at all easy to code for, or even getting the stuff to compile right. Its also less mature and less supported than CUDA, making.. coding for CUDA arguably actually easier overall. For all the hammering away at some of the problems I was facing, I could only get around half the benchmarks to run on AVX512 intel cpus without crashing. Some of them still produced unexpected outputs. Anyways, verdict is that, while GPUs and dedicated hardware still easily outpace CPUs here, the 2990WX isnt half bad at it either if you cant otherwise use GPUs.
Games: it manages a solid 60fps at 4k using dual 1080tis. Cant ask for much more than that. I'm an artist and programmer who games on the side, I like high resolutions, and I dont really play anything that greatly benefits from high fps - it just makes it nicer to watch but provides no tangible benefits for what I play.
If you have workloads that benefit from this CPU, you or at least your business probably makes enough money to get this CPU, and in the arena that this CPU fights in, where its battling parts that range from $500 to $10000, its one of the best values there is for workloads its best at.
If you also want to game on it, well, its perfectly fine if you just want 60hz or even 90hz (vr), with dynamic local mode bringing it closer to where the 2950X and 2700X sit. Its never going to be a top tier champ of the high refresh rate gaming arena, but thats not what its value proposition is for anyways, theres cpus better at that for much cheaper (2700X, 8700K, and just forget the 9900K even exists since the 8700K is just a few percent slower while not running into the same power/thermal issues as the 9900K - which by the way, actually can use more power (I've seen people reporting up to 265W on the package) than even the 2990WX (peak power draw I've observed is 248W) does at stock settings on a good board under all-core loads, since MCE is a stock setting enabled by default on a lot of those boards that can actually supply it enough power without the VRMs throttling or overheating).
